Nvidia open-sources Flora, a data engine that labels robot training video for 72 cents an hour
Nvidia compares that with $88 an hour for GPT-6 Astra, though The Decoder says the gap is narrower than the price suggests.
Nvidia's GEAR research lab has open-sourced Flora, a data engine that annotates video used to train robot AI models. The models are available on Hugging Face. Nvidia says robot models need three kinds of labels: descriptions of which action happens when, used to select training data; step-by-step ratings of how well a robot is progressing, used for trial-and-error learning; and judgments on whether a task was completed, used to monitor robots in deployment. Most existing tools provide only one of these. Large language models such as GPT-6 Astra can produce several, but they cost too much at the volumes involved. Flora splits videos into time segments, describes each action, and evaluates both the execution and the outcome of an episode. Nvidia puts the cost at 72 cents per hour of video, compared with $88 for GPT-6 Astra. The Decoder cautions that the comparison is less lopsided than that price gap implies. Nvidia says Flora already runs during training of its GR00T robot models, annotating up to 200,000 hours of video a day across 512 GB300 GPUs. Cheaper labeling could lower one of the main costs for groups building robot foundation models.